Anthropic’s Applied AI team published a piece on 22 October that cuts against the grain. The argument: plenty of companies build too many subagents, and they do better folding routers, planners and domain specialists back into the main loop.
Where the subagents come from
The pattern repeats everywhere. A company wants several products behind one interface. Each capability belongs to a different team, so every team builds its own agent and a platform team puts an orchestrator on top. It looks like the simplest answer and brings four problems, which Anthropic names outright.
Context: somebody has to decide what gets passed between agents, and detail goes missing in the handoffs. Sprawl: every team ships its subagent, and you end up with hundreds of overlapping ones, no enforced standards, no owner. The Frankenprompt: every team adds a sentence to the system prompt until the rules contradict each other and no consistent voice survives. Evals: everyone optimises for their own subtask, nobody maintains the whole set, and leadership decides on watered-down scores.
Then there is speed. One customer measured its multi-agent harness against a single agent loop and saw the same work run about twice as fast.
What works instead
Anthropic’s proposal is progressive disclosure. Context arrives through skills, tools through tool search. The agent sees short descriptions first and loads the full thing only when it needs it. Across comparisons at several enterprise deployments, a single agent with skills frequently beat both alternatives, the one-prompt-for-everything design and the subagent design, often at lower cost and latency per task.
One example from the post: a customer-facing agent with a dozen specialist subagents looked tidy on the whiteboard. In practice, keeping a consistent on-brand experience across all of them was hard, and the extra latency hurt the user experience.
When subagents stay
Anthropic is not writing them off. Three cases still hold: when a single loop shows measurable degradation from context bloat, when tasks need to run in parallel with fresh context windows that get thrown away afterwards, and when one agent struggles to pick the right tool or the system prompt contains conflicting instructions. For verification, parallel work and separate permissions, subagents remain the right tool, and Anthropic uses them in its own systems.
If you want to check what belongs merged in your own setup, there is a migration kit with a guided workflow and a demo against a mock system.
An org chart is not an architecture
The line that sticks is the aside about Conway’s law: too often each team and each problem type gets its own subagent, and the architecture becomes a copy of the org chart. That is an organisational argument, not a technical one, which is exactly why no better framework fixes it. Often the structure was doing work the model no longer needs done for it: one clear system prompt, four or five thin data tools, firm guardrails on the output.